If you upgraded to a paid Claude plan just to use the new Fable 5 and Mythos 5 AI models, you’re going to want to sit down for bad news. The US government has issued an emergency national security directive forcing Anthropic to suspend access due to concerns over a security vulnerability.
According to the report, the issue centers on a jailbreak exploit that allows the models to be manipulated into scanning codebases for software weaknesses. Regulators allegedly treated it as a national security risk and ordered access restrictions for foreign users. Anthropic has now disabled access for all users globally, including paying subscribers and even US customers.
The US government fears Claude jailbreaking
Last month, Anthropic introduced updates to its Claude lineup, upgrading its flagship model and expanding how users can control its performance. The new version focuses on reduced hallucinations and better alignment with user intent.
It also introduces effort controls that allow users to choose how much computing power the model uses for a given task, along with early work on more advanced agent-style workflows that can break complex tasks into multiple parallel processes. Both models cost $10 per 1 million input tokens and $50 per 1 million output tokens.
However, days ago, they released an official statement explaining that they're now disabling access to Fable 5 and Mythos 5 after receiving a government order.
We received the directive from the government today at 5:21pm (ET). The letter did not provide specific details of its national security concern. Our understanding is that the government believes it has become aware of a method of bypassing, or “jailbreaking” Fable 5.
We reviewed a demonstration of this specific technique being used to identify a small number of previously known, minor vulnerabilities. These vulnerabilities all appear relatively simple, and we have found that other publicly-available models are able to discover them as well without requiring a bypass.
Anthropic
Although the government didn’t give them a detailed explanation, the directive is to block permission for any foreign national and even employees working inside the company. If Anthropic’s guess is right, their aim is to prevent jailbreaking.
What is jailbreaking?
Jailbreaking, in this context, refers to someone tricking AI models into bypassing its safety rules and performing cybersecurity-related tasks, specifically scanning a codebase for vulnerabilities.
Although it's impossible to selectively enforce the US government's new changes, Anthropic is forced to comply. They reviewed the jailbreak and found that it only exposes minor, already-known issues. They also argue that other AI models can already do the same thing without needing any bypass, and defend Fable 5’s creation process.
According to them, their model is trained to refuse harmful requests and additional filters sit around it to catch risky inputs and outputs. It also monitors systems looking for patterns of misuse over time.
To the government’s defence, Anthropic has spent years publicly flagging that advanced models will never be fully jailbreak-proof and that some level of misuse is unavoidable. If you keep crying wolf long enough, one is bound to show up.
